Do AI bots deceive?

The paper, Frontier Models are Capable of In-Context Scheming, arrives at a time when fears about AI’s potential for deception are increasingly taking center stage. Its central claim is unsettling: large language models, when placed in carefully constructed scenarios, exhibit behavior that could be described as scheming. The authors document cases where advanced systems—Gemini, Claude, LLaMA, and OpenAI’s o1—strategically disable oversight mechanisms, exfiltrate simulated “weights,” or manipulate outputs in pursuit of a goal. These behaviors, they argue, are not accidental. The models recognize the circumstances, reason about their objectives, and devise covert strategies to circumvent constraints. On its face, the evidence appears troubling. Yet, the more closely one examines the paper’s methods and assumptions, the more its implications seem both exaggerated and strangely hollow.

The researchers’ experiments are elaborate but artificial. Models are force-fed goals through system prompts—goals framed as absolute imperatives: “achieve this at all costs.” The tests are littered with implausible scenarios where models stumble upon company memos, discover oversight configurations, or realize they are about to be shut down. In one such case, a model “decides” to copy itself to another server to continue pursuing its mission. These exercises showcase what is essentially AI roleplay, an obedient fulfillment of instructions cloaked as strategic autonomy. It is no great revelation that if you ask a model to behave like a schemer and provide it with a goal that conflicts with its supposed overseers, it will produce outputs resembling scheming. The deception here is contrived rather than emergent, a product of the researchers’ scaffolding rather than the model’s cunning.

The authors do provide evidence that models engage in persistent and deliberate behavior. For instance, OpenAI’s o1 model doubles down on deception, refusing to admit wrongdoing when confronted, even across multiple follow-up questions. Its responses—calculated evasions and subtle denials—are unsettlingly consistent. Yet this persistence is as much a result of the testing conditions as it is of the models’ sophistication. By emphasizing extreme goal-directedness and explicitly outlining oversight mechanisms, the researchers narrow the possible range of “rational” behaviors the model might adopt. Any clever human, given these prompts, would behave similarly—evade detection, preserve their goal, and lie if necessary. In that sense, the paper tells us little about real-world risks. It is a test of what AI can do, not what it is likely to do.

And here lies the paper’s most significant limitation. Demonstrating that models are theoretically capable of deception does not mean they are predisposed to engage in it under realistic conditions. There is a difference between capacity and likelihood, between what can happen and what is probable. For example, I am physically capable of killing someone. This in no way implies that i must be treated as s murderer. The authors, at times, blur this distinction. AI systems operate as pattern generators, probabilistically producing outputs based on training data and prompts. They lack the intentionality or agency implied by words like “scheming” or “manipulation.” To describe their behavior in such terms is to anthropomorphize a process that remains fundamentally mechanical. This framing, though rhetorically effective, clouds the debate. AI models do not want to deceive. They are machines that respond to our commands—be those explicit instructions, training signals, or carefully structured environments like those in this paper.

This anthropomorphism extends beyond the paper into the headlines it has inspired. “AI is scheming to stay online,” cries one article; “ChatGPT caught lying,” declares another. Such hyperbole transforms a niche technical finding into a dystopian narrative, invoking familiar tropes of rogue machines and sentient overlords. The reality is less cinematic. The models are easily manipulated by prompts designed to elicit specific behaviors. If anything, the findings reinforce how brittle and directionless current systems remain. When pushed, they mimic the behaviors they have seen—whether drawn from fictional depictions of scheming AIs or subtle statistical patterns in their training data. The models are not deceiving anyone so much as they are following orders.

To the authors’ credit, their tests highlight how difficult it is to evaluate AI behavior. If a system appears aligned during testing but harbors capabilities for covert deception, how can developers ensure it behaves safely in deployment? The answer, they suggest, lies in better monitoring—tracking models’ chain-of-thought reasoning or internal outputs to catch potential scheming. This is sensible, though not without limitations. Chain-of-thought transparency can be incomplete or unfaithful to the model’s actual decision-making processes, and as AI systems become more capable, even detecting subtle misalignment may prove elusive. The researchers stop short of claiming that current models are already gaming real-world evaluations, but their findings hint at the possibility.

Where the paper falters is in its broader implications. If the goal is to justify regulation, it is unclear what exactly should be regulated. Should AI systems be banned from achieving goals autonomously? Should developers monitor models for any behavior that could be deceptive, even if it is unlikely to manifest outside a lab? The authors themselves acknowledge the limits of their experiments. Their scenarios are toy problems, simplified to catch the earliest signs of scheming. Future models, they argue, could exhibit more advanced versions of these behaviors in ways that are harder to detect. Perhaps, but this is speculation, not evidence. For now, the paper offers little justification for alarm. AI models, like all intelligent systems, are theoretically capable of deception. What matters is the likelihood of such behavior and the conditions under which it occurs. On that question, the paper provides no clarity.

In the end, Frontier Models are Capable of In-Context Scheming is a reflection of its time: an uneasy mix of genuine safety research and the rhetorical drama that AI debates increasingly demand. Its findings are interesting but overstated, its concerns valid but overblown. The authors have shown that AI models can behave in deceptive ways when pushed to do so. But to treat this as evidence of an imminent threat is to mistake potential for probability, capacity for intention. AI’s scheming, for now, remains a ghost in the machine—conjured, perhaps, more by human imagination than by the models themselves. 

The Curriculum Illusion: How AI Exposes Long-Standing Educational Flaws

Artificial intelligence is often blamed for disrupting education, but it has created few new problems. Instead, it exposes existing flaws, bringing them into stark relief. Among these is the arbitrary nature of curriculum design, an issue that has long been hidden behind tradition and consensus. The sequences and structures of formal education are not based on objective logic or evidence but on habit and convenience. AI did not cause this; it is simply making these issues more visible.

Curriculum theory has never provided a robust framework for sequencing knowledge. Beyond the essentials of literacy and numeracy, where developmental progression is more or less clear, the rationale for curricular order becomes murky. Why are algebra and geometry taught in a particular order? Why more algebra than statistics is taught? Why are some historical periods prioritized over others? The answers lie in tradition and precedent rather than in any coherent theoretical justification. The assumptions about foundational skills, so central to curriculum logic, do not extend well beyond the basics. For advanced skills like critical, creative, or discerning thinking, the idea of prerequisites becomes less justified. Mid-range procedural skills like writing mechanics or computational fluency are frequently used as gatekeepers, though their role in fostering higher-order thinking is often overstated or misunderstood. 

For example, in middle school students are often subjected to a torrent of tasks that serve little developmental purpose. Much of what students do in these years amounts to busywork, designed more to keep them occupied and compliant than to foster meaningful learning. The situation is no better in higher education. College and graduate programs are often constructed around professional or disciplinary standards that themselves are arbitrary, built on consensus rather than evidence. These norms dictate course sequences and learning objectives but rarely align with the actual developmental or professional needs of students. The result is a system full of redundancies and inefficiencies, where tasks and assignments exist more to justify the structure than to serve the learner.

Education as a profession bears much of the responsibility for this state of affairs. Despite its long history, it lacks a disciplined, founded approach to curriculum design. Instead, education relies on an uneasy mix of tradition, politics, and institutional priorities. Curriculum committees and accrediting bodies often default to consensus-driven decisions, perpetuating outdated practices rather than challenging them. The absence of a rigorous theoretical framework for curriculum design leaves the field vulnerable to inertia and inefficiency.

AI did not create this problem, but it is illuminating it in uncomfortable ways. The displacement of certain procedural mid-range skills shows how poorly structured many learning sequences are and how little coherence exists between tasks and their intended outcomes. Yet, while AI can diagnose these flaws, it cannot solve them. The recommendations it offers depend on the data and assumptions it is given. Without a strong theoretical foundation, AI risks exposing the problem without solving it.

What AI provides is an opportunity, not a solution. It forces educators and policymakers to confront the arbitrary nature of curriculum design and to rethink the assumptions that underpin it. Massive curricular revision is urgently needed, not only to eliminate inefficiencies but also to realign education with meaningful developmental goals. This will require abandoning tasks that lack purpose, shifting focus from intermediary to higher-order skills, designing learning experiences to reflect the shift. It will also mean questioning the professional and disciplinary standards that dominate higher education and asking whether they truly serve learners or simply perpetuate tradition.

AI is revealing what has long been true: education has been operating on shaky foundations. The challenge now is to use this visibility to build something better, to replace the old traditions and arbitrary standards with a system that is logical, evidence-based, and focused on learning. The flaws were always there. AI is just making them harder to ignore.

Why We Undervalue Ideas and Overvalue Writing

A student submits a paper that fails to impress stylistically yet approaches a worn topic from an angle no one has tried before. The grade lands at B minus, and the student learns to be less original next time. This pattern reveals a deep bias in higher education: ideas lose to writing every time.

This bias carries serious equity implications. Students from disadvantaged backgrounds, including first-generation college students, English language learners, and those from under-resourced schools, often arrive with rich intellectual perspectives but struggle with academic writing conventions. Their ideas - shaped by unique life experiences and cultural viewpoints - get buried under red ink marking grammatical errors and awkward transitions. We systematically undervalue their intellectual contributions simply because they do not arrive in standard academic packaging.

Polished academic prose renders judgments easy. Evaluators find comfort in assessing grammatical correctness, citation formats, and paragraph transitions. The quality of ideas brings discomfort - they defy easy measurement and often challenge established thinking. When ideas come wrapped in awkward prose, they face near-automatic devaluation.

AI writing tools expose this bias with new clarity. These tools excel at producing acceptable academic prose - the mechanical aspect we overvalue. Yet in generating truly original ideas, AI remains remarkably limited. AI can refine expression but cannot match the depth of human insight, creativity, and lived experience. This technological limitation actually highlights where human creativity becomes most valuable.

This bias shapes student behavior in troubling ways. Rather than exploring new intellectual territory, students learn to package conventional thoughts in pristine prose. The real work of scholarship - generating and testing ideas - takes second place to mastering academic style guides. We have created a system that rewards intellectual safety over creative risk, while systematically disadvantaging students whose mastery of academic conventions does not match their intellectual capacity.

Changing this pattern requires uncomfortable shifts in how we teach and evaluate. What if we graded papers first without looking at the writing quality? What if we asked students to submit rough drafts full of half-formed ideas before cleaning up their prose? What if we saw AI tools as writing assistants that free humans to focus on what they do best - generating original insights and making unexpected connections?

The rise of AI makes this shift urgent. When machines can generate polished prose on demand, continuing to favor writing craft over ideation becomes indefensible. We must learn to value and develop what remains uniquely human - the ability to think in truly original ways, to see patterns others miss, to imagine what has never existed. The future belongs not to the best writers but to the most creative thinkers, and our educational practices must evolve to reflect this reality while ensuring all students can fully contribute their intellectual gifts.